RESUMO
Evapotranspiration (ET) is a fundamental flux in land surface hydrothermal process. Because of the differences in basic concepts, assumptions, application scales, different models have induced varying uncertainties to the estimation and simulation of evapotranspiration. With the Three-River-Source National Park as an example, we used the Bayesian model averaging (BMA) method to integrate the ET estimations from five models of PT-JPL, ARTS-GIMMS3, ARTS-MODIS, MODIS global evapotranspiration product (MOD16), and SSEBop, and tried to improve the estimating accuracy of evapotranspiration. The results showed that the five models could well capture the seasonal variations in evapotranspiration at Haibei Flux Station, with an explanation range of 64%-86% variability in the observed ET, and a root means square deviation (RMSD) ranged from 0.47 mm·(8 d)-1 to 0.76 mm·(8 d)-1. BMA-based ET greatly improved its explanation to 89% and decreased the RMSD to 0.43 mm·(8 d)-1. The Three-River-Source National Park experienced an overall insignificant increasing trend in its inter-annual ET from 2003 to 2015. At the regional scale, the effects of temperature and precipitation on evapotranspiration were not significant, but were significant in the Yangtze River Source Park. Temperature and precipitation had positive impacts on evapotranspiration. The evapotranspiration showed different trends due to the geographi-cal differences between parks. This study provided a method reference for other multi-source data integration analysis. The integrated evapotranspiration data could effectively reduce the uncertainty of the original models and provide a more accurate data basis for the study of regional water heat change, which is of great significance to better understand water cycle under climate changes.